Comparative Analysis of Elliptic Envelope, Isolation Forest, One-Class SVM, and Local Outlier Factor in Detecting Earthquakes with Status Anomaly using Outlier

Abstract

In Indonesia, large earthquakes happen frequently. Due to the location of Indonesia, over three active tectonic plates, fractures and trenches have formed on both the land and ocean floor. Over 1,250 large earthquake occurrences have occurred in Indonesia during the past 120 years, beginning in 1900. Based on data from the BMKG (Indonesian non-departmental government organization for meteorology, climatology, and geophysics) official website, up to 12,351 earthquakes occurred in Indonesia in 2021. Not every earthquake was fatal; many of them were considered common. In this work, the author compares Elliptic Envelope, Isolation Forest, One-Class SVM, and Local Outlier Factor. 4 types of anomaly detection algorithms to locate earthquakes whose status is abnormal or out of the ordinary utilizing outliers. The application of anomaly detection for each algorithm will be tested in two ways; the first test will only use the default parameters, and the second test will use the parameter (Tuning Hyperparameters) at the contamination/nu parameter of 0.025 or 2.5%. Contamination/nu is a value that indicates the amount of data considered an anomaly—Accuracy, Recall, Precision, F1-Score, AUC, and Specificity. Experiments have shown that Isolation Forest has a higher value than all other experiments in detecting anomalous status in earthquakes. In comparing performance using default parameters, Isolation Forest had a precision of 13.02%, recall (sensitivity) of 99.35%, specificity of 82.80%, accuracy of 83.24%, AUC of 91.08%, and F1-score of 23.03%. After tuning the hyperparameters, the performance of each model increased, with Isolation Forest had a precision of 37.21%, recall (sensitivity) of 36.85%, specificity of 98.41%, accuracy of 98.42%, AUC of 67.63%, and F1-score of 37.03%.